Gemma 4 QAT 31B responds better to KV cache quantization

Benchmarks show Gemma 4 QAT 31B improves performence with KV cache quantization, building on earlier findings for the 9B model. The 31B variant achieves better results under reduced cache precision.